1#ifndef CEPH_CRUSH_CRUSH_H 2#define CEPH_CRUSH_CRUSH_H 3 4#include <linux/types.h> 5 6/* 7 * CRUSH is a pseudo-random data distribution algorithm that 8 * efficiently distributes input values (typically, data objects) 9 * across a heterogeneous, structured storage cluster. 10 * 11 * The algorithm was originally described in detail in this paper 12 * (although the algorithm has evolved somewhat since then): 13 * 14 * http://www.ssrc.ucsc.edu/Papers/weil-sc06.pdf 15 * 16 * LGPL2 17 */ 18 19 20#define CRUSH_MAGIC 0x00010000ul /* for detecting algorithm revisions */ 21 22#define CRUSH_MAX_DEPTH 10 /* max crush hierarchy depth */ 23 24 25#define CRUSH_ITEM_UNDEF 0x7ffffffe /* undefined result (internal use only) */ 26#define CRUSH_ITEM_NONE 0x7fffffff /* no result */ 27 28/* 29 * CRUSH uses user-defined "rules" to describe how inputs should be 30 * mapped to devices. A rule consists of sequence of steps to perform 31 * to generate the set of output devices. 32 */ 33struct crush_rule_step { 34 __u32 op; 35 __s32 arg1; 36 __s32 arg2; 37}; 38 39/* step op codes */ 40enum { 41 CRUSH_RULE_NOOP = 0, 42 CRUSH_RULE_TAKE = 1, /* arg1 = value to start with */ 43 CRUSH_RULE_CHOOSE_FIRSTN = 2, /* arg1 = num items to pick */ 44 /* arg2 = type */ 45 CRUSH_RULE_CHOOSE_INDEP = 3, /* same */ 46 CRUSH_RULE_EMIT = 4, /* no args */ 47 CRUSH_RULE_CHOOSELEAF_FIRSTN = 6, 48 CRUSH_RULE_CHOOSELEAF_INDEP = 7, 49 50 CRUSH_RULE_SET_CHOOSE_TRIES = 8, /* override choose_total_tries */ 51 CRUSH_RULE_SET_CHOOSELEAF_TRIES = 9, /* override chooseleaf_descend_once */ 52 CRUSH_RULE_SET_CHOOSE_LOCAL_TRIES = 10, 53 CRUSH_RULE_SET_CHOOSE_LOCAL_FALLBACK_TRIES = 11, 54 CRUSH_RULE_SET_CHOOSELEAF_VARY_R = 12 55}; 56 57/* 58 * for specifying choose num (arg1) relative to the max parameter 59 * passed to do_rule 60 */ 61#define CRUSH_CHOOSE_N 0 62#define CRUSH_CHOOSE_N_MINUS(x) (-(x)) 63 64/* 65 * The rule mask is used to describe what the rule is intended for. 66 * Given a ruleset and size of output set, we search through the 67 * rule list for a matching rule_mask. 68 */ 69struct crush_rule_mask { 70 __u8 ruleset; 71 __u8 type; 72 __u8 min_size; 73 __u8 max_size; 74}; 75 76struct crush_rule { 77 __u32 len; 78 struct crush_rule_mask mask; 79 struct crush_rule_step steps[0]; 80}; 81 82#define crush_rule_size(len) (sizeof(struct crush_rule) + \ 83 (len)*sizeof(struct crush_rule_step)) 84 85 86 87/* 88 * A bucket is a named container of other items (either devices or 89 * other buckets). Items within a bucket are chosen using one of a 90 * few different algorithms. The table summarizes how the speed of 91 * each option measures up against mapping stability when items are 92 * added or removed. 93 * 94 * Bucket Alg Speed Additions Removals 95 * ------------------------------------------------ 96 * uniform O(1) poor poor 97 * list O(n) optimal poor 98 * tree O(log n) good good 99 * straw O(n) better better 100 * straw2 O(n) optimal optimal 101 */ 102enum { 103 CRUSH_BUCKET_UNIFORM = 1, 104 CRUSH_BUCKET_LIST = 2, 105 CRUSH_BUCKET_TREE = 3, 106 CRUSH_BUCKET_STRAW = 4, 107 CRUSH_BUCKET_STRAW2 = 5, 108}; 109extern const char *crush_bucket_alg_name(int alg); 110 111struct crush_bucket { 112 __s32 id; /* this'll be negative */ 113 __u16 type; /* non-zero; type=0 is reserved for devices */ 114 __u8 alg; /* one of CRUSH_BUCKET_* */ 115 __u8 hash; /* which hash function to use, CRUSH_HASH_* */ 116 __u32 weight; /* 16-bit fixed point */ 117 __u32 size; /* num items */ 118 __s32 *items; 119 120 /* 121 * cached random permutation: used for uniform bucket and for 122 * the linear search fallback for the other bucket types. 123 */ 124 __u32 perm_x; /* @x for which *perm is defined */ 125 __u32 perm_n; /* num elements of *perm that are permuted/defined */ 126 __u32 *perm; 127}; 128 129struct crush_bucket_uniform { 130 struct crush_bucket h; 131 __u32 item_weight; /* 16-bit fixed point; all items equally weighted */ 132}; 133 134struct crush_bucket_list { 135 struct crush_bucket h; 136 __u32 *item_weights; /* 16-bit fixed point */ 137 __u32 *sum_weights; /* 16-bit fixed point. element i is sum 138 of weights 0..i, inclusive */ 139}; 140 141struct crush_bucket_tree { 142 struct crush_bucket h; /* note: h.size is _tree_ size, not number of 143 actual items */ 144 __u8 num_nodes; 145 __u32 *node_weights; 146}; 147 148struct crush_bucket_straw { 149 struct crush_bucket h; 150 __u32 *item_weights; /* 16-bit fixed point */ 151 __u32 *straws; /* 16-bit fixed point */ 152}; 153 154struct crush_bucket_straw2 { 155 struct crush_bucket h; 156 __u32 *item_weights; /* 16-bit fixed point */ 157}; 158 159 160 161/* 162 * CRUSH map includes all buckets, rules, etc. 163 */ 164struct crush_map { 165 struct crush_bucket **buckets; 166 struct crush_rule **rules; 167 168 __s32 max_buckets; 169 __u32 max_rules; 170 __s32 max_devices; 171 172 /* choose local retries before re-descent */ 173 __u32 choose_local_tries; 174 /* choose local attempts using a fallback permutation before 175 * re-descent */ 176 __u32 choose_local_fallback_tries; 177 /* choose attempts before giving up */ 178 __u32 choose_total_tries; 179 /* attempt chooseleaf inner descent once for firstn mode; on 180 * reject retry outer descent. Note that this does *not* 181 * apply to a collision: in that case we will retry as we used 182 * to. */ 183 __u32 chooseleaf_descend_once; 184 185 /* if non-zero, feed r into chooseleaf, bit-shifted right by (r-1) 186 * bits. a value of 1 is best for new clusters. for legacy clusters 187 * that want to limit reshuffling, a value of 3 or 4 will make the 188 * mappings line up a bit better with previous mappings. */ 189 __u8 chooseleaf_vary_r; 190}; 191 192 193/* crush.c */ 194extern int crush_get_bucket_item_weight(const struct crush_bucket *b, int pos); 195extern void crush_destroy_bucket_uniform(struct crush_bucket_uniform *b); 196extern void crush_destroy_bucket_list(struct crush_bucket_list *b); 197extern void crush_destroy_bucket_tree(struct crush_bucket_tree *b); 198extern void crush_destroy_bucket_straw(struct crush_bucket_straw *b); 199extern void crush_destroy_bucket_straw2(struct crush_bucket_straw2 *b); 200extern void crush_destroy_bucket(struct crush_bucket *b); 201extern void crush_destroy_rule(struct crush_rule *r); 202extern void crush_destroy(struct crush_map *map); 203 204static inline int crush_calc_tree_node(int i) 205{ 206 return ((i+1) << 1)-1; 207} 208 209#endif 210